***********************************************************************
ADVISORY: TCP packet fragment attacks against firewalls and filters
  System: TCP/IP networks  
  Source: http://all.net, Dr. Frederick B. Cohen
***********************************************************************

Packet Fragmentation Attacks

Introduction to Packet Fragmentation

Packet fragmentation is the part of the Internet Protocol (IP) suite of
networking protocols that assures that IP datagrams can flow through any
other sort of network. (For details, see Internet Request For Comments 791
(rfc791) and are available and searchable in electronic form from Info-Sec
heaven on the World-Wide-Web at http://all.net, through gopher service at
all.net, or by ftp service from rs.internic.net.) Fragmentation works by
allowing datagrams created as a single packet to be split into many smaller
packets for transmission and reassembled at the receiving host.

Packet fragmentation is necessary because underlying the IP protocol, other
physical and or logical protocols are used to transport packets through
networks. A good example of this phenomena is on the difference between
Ethernet packets (which are limited to 1024 bytes), ATM packets (which are
limited to 56 bytes), and IP packets which have variable sizes up to about
1/2 million bytes in length.

The only exception to this rule is in the case of an internet datagram
marked don't fragment . Any internet datagram marked in this way is
supposed to not be fragmented under any circumstances. If internet
datagrams marked don't fragment cannot be delivered to their destination
without being fragmented, they are supposed to be discarded instead. Of
course, this rule doesn't have to be obeyed by the IP software actually
processing packets, but it is supposed to be.

How Packet Reassembly Attacks Work

The packet fragmentation mechanism leads to attacks that bypass many
current Internet firewalls, but the reason these attacks work is not
because of the way fragmentation is done, but rather because of the way
datagrams are reassembled.

Datagrams are supposed to be fragmented into packets that leave the header
portion of the packet intact except for the modification of the fragmented
packet bit and the filling in of an offset field in the IP header that
indicates at which byte in the whole datagram the current packet is
supposed to start. In reassembly, the IP reassembler creates a temporary
packet with the fragmented part of the datagram in place and adds incoming
fragments by placing their data fields at the specified offsets within the
datagram being reassembled. Once the whole datagram is reassembled, it is
processed as if it came in as a single packet.

According to the IP specification, fragmented packets are to be reassembled
at the receiving host. This presumably means that they are not supposed to
be reassembled at intermediate sites such as firewalls or routers. This
decision was made presumably to prevent repeated reassembly and
refragmentation in intermediate networks. When routers and firewalls
followed the rules, they found a peculiar problem.

The way firewalls and routers block specific services (such as telnet )
while allowing other services (such as the world wide web http service) is
by looking into the IP packet to determine which Transfer Control Protocol
(TCP) port is being used. If the port corresponds to 80, the datagram is
destined for http service, while port 23 is used for telnet . In normal
datagrams, this works fine. But suppose we didn't follow the rules for
fragmentation and created improper fragmented packets? Here's what one
attacker did:

   *  Create an initial packet which claims to be the first fragment of a
     multi-packet datagram. Specify TCP port 80 in the TCP header so it
     looks like a datagram going to http service, which is allowed to pass
     the firewall.
   *  The firewall passes the packet to the host under attack and passes
     subsequent packet fragments in order to allow the destination host to
     reassemble the packet.
   *  One of the subsequent packets has an offset of 0 which causes the
     reassembler to overwrite the initial part of the IP packet. This is
     the part of the IP packet that specifies the TCP port. The attacker
     overwrites the IP port number which was originally 80 with a new port
     number such as 23, and is now granted telnet access to the host under
     attack despite the firewall that is supposed to block the service.